relstorage_packer 2.1

Overview

This script works also for very large Relstorage ZODBs with several million objects. The original pack script took several days and consumed lots of RAM. So there was need to accelerate the process of packing.

This script does not consume relevant amounts of RAM, runs much faster than the original. Where the old took 3.5 days only for analysis it takes now about 6 hours. On subsequent runs it only processes changes after last run: it considers only transactions newer than last processed transaction of the prior run.

At time of writing processing 44mio objects takes initially about 3-6h depending on hardware and configuration of Postgresql.

The script creates an inverse object graph, this takes little extra space in DB.

Limitations

At time of development the critical production environment was a postgresql database running relstorage with blobs stored on a fileserver in history free mode. So this is implemented.

I’am sure its easily possible to make this work on MySQL and Oracle too. Also considering blobs inside DB is for sure possible.

I’am not sure if this way of cleanup makes sense for non-history-free mode. At least it needs a lot of love and understanding of ZODB to refactor and implement.

Contributions are welcome!

Usage

Create a configuration file. It is the same as used in classical pack script deployed with Relstorage:

<relstorage>
    create-schema false
    keep-history false
    shared-blob-dir true
    blob-dir var/blobstorage
    commit-lock-timeout 600
    <postgresql>
        dsn dbname='test_site' host='127.0.0.1' user='zodb' password='secret'
    </postgresql>
</relstorage>

After installation a script relstorage_pack is available:

Usage: relstorage_pack config_file

Fast ZODB Relstorage Packer for history free PostgreSQL

Options:
  -h, --help     show this help message and exit
  -i, --init     Removes all reference counts and starts from scratch.
  -v, --verbose  More verbose output, includes debug messages.

When running first time with your database pass --init as parameter. This drops and recreates the packing table.

Using integrated buildout and testing

Testing this code i not easy and writing good tests is a task to be done. At the moment you can try the code bu running a postgres database on localhost (unless you want to change buildout.cfg). Then run as database-user (named postgres on debian) the commands:

psql -c "CREATE USER zope WITH PASSWORD 'secret';"
psql -c "CREATE DATABASE relstorage_packer_test OWNER zope;"
psql -c "REVOKE connect ON DATABASE relstorage_packer_test FROM PUBLIC;"
psql -c "GRANT connect ON DATABASE relstorage_packer_test TO zope;"

Next (because of my laziness) run ./bin/instance start which starts a Plone. Add a Plone Site, add and delete some content to fill the database with something to pack.

Next run the packer.

If you dont like this: pull requests are always welcome.

2.1 (2014-02-19)

• refactored logging, because it was always in my way when changing other parts of the code. [jensens, 2014-02-19]

• after long running connections postgresql takes lots of RAM. So we reconnect every 5000 cycles (TID analyzing or ZOID removal). [jensens, 2014-02-18]

• we had a whole bunch of idle in transaction (aborted) postgres processes running after packing. This resulted in in an OperationalError: out of shared memory HINT: You might need to increase max_pred_locks_per_transaction. Error. As a result I refactored the transaction handling and rollback and use explicit commit instead of using the relstorage storage._with_store. Now this part is very controlled and not the source of hanging connections w/o rollback. [jensens, 2014-02-18]

2.0.2 (2014-02-05)

• also support storages w/o blobstorage [jensens, 2014-02-05]

2.0.1 (2014-02-03)

• unlock in a finally to really unlock an failure [jensens, 2014-02-03]

• use non-zero exit code if lock could not be aquired. [saily]

2.0

• refactored the way of collecting and using the reference counts. faster now. [jensens, 2014-01-11]

1.0

• started package [jensens, 2013-11-23]